Date: Wed, 29 May 2002 21:35:18 -0500 From: Dan Nelson <dnelson@allantgroup.com> To: Gary Stanley <gary@outloud.org> Cc: freebsd-hackers@FreeBSD.ORG Subject: Re: mbuf problems on 4.5/4.6-RC2 Message-ID: <20020530023518.GB78068@dan.emsphone.com> In-Reply-To: <5.1.1.2.2.20020529204459.00b39c48@208.141.46.254> References: <5.1.1.2.2.20020529204459.00b39c48@208.141.46.254>
next in thread | previous in thread | raw e-mail | index | archive | help
In the last episode (May 29), Gary Stanley said: > We have a webserver setup with 1 Intel Pro/1000 Gigabit Ethernet > card. The old server was using around 95mbit/s of traffic or so > (sustained rate). We decided to replace that card/machine with > something that could handle the high network load. The question would > be, What's a good size for NMBCLUSTERS? I want to be -sure- the box > does not run out of mbufs during the production phase. The old > machine was on a fxp0 driver, mbufs set to 65k due to network > overhead. Are there limits to mbufs, as in a size limit? mbufs are more related to the number of active sockets and the window size than total throughput. I've got two boxes with gigabit NICs that do up to 40MBytes/sec but have never peaked above 3k mbufs, since the traffic is over maybe 10 sockets. If you have lots of connections, you might want to shrink your window size to compensate. Assuming you're sending lots of data to clients whose individual throughput is much less than yours, dropping your send window to 16k (from 32k) and your receive window to 8k (from 64k) will cut your mbuf usage to 1/4 its previous value. 64k mbufs is probably too high, actually. The tuning(7) manpage goes into a bit more detail. If you are still having problems, try the -net mailinglist. -- Dan Nelson dnelson@allantgroup.com To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20020530023518.GB78068>